- Nvidia upgraded its Spectrum-X Ethernet platform to allow AI workloads to more easily scale across geographically different data center sites
- No hardware changes are needed, Spectrum-XGS is all about algorithm-based improvements
- Nvidia also highlighted progress it is making with its Dynamo disaggregated serving tools for inference workloads
Forget scale up and scale out. Scale across is the hot new trend in data center networking.
Nvidia on Friday unveiled an upgraded version of its Spectrum-X Ethernet platform – called Spectrum-XGS – designed to enable AI demand to be shared across multiple data centers.
You remember when the Power Rangers assembled their individual vehicles into one giant battle robot? Yeah, it’s kind of like that and Spectrum-XGS is the connective tissue holding it all together and instead of robots it’s data centers (ok, so maybe still robots).
Gilad Shainer, Nvidia’s SVP of networking, told journalists on a call that there’s no new hardware involved in Spectrum-XGS. Rather, he said, Nvidia’s team “created new algorithms that enable the effectiveness of moving data across longer distances between sites.”
How much does the upgrade matter? Well, Nvidia claimed that Spectrum-XGS increases multi-sight performance of its Collective Communications Library by 1.9x and provides automatic load balancing adjustments. The sum is more predictable and performant multi-GPU and multi-node communication within geographically distributed AI clusters.
The tweak is notably given the growing momentum behind Nvidia's Spectrum-X platform. The company doesn't report its fiscal Q2 earnings until next week, but company executives said on its FQ1 call that Spectrum-X was already "annualizing over $8 billion in revenue," which factors out to roughly $2 billion in revenue per quarter. Spectrum-X customers include Microsoft, Oracle and CoreWeave, with Google Cloud and Meta the latest to adopt it.
Distributed AI
The distribution of AI workloads seems to be a theme for Nvidia. The company on the call also talked up progress it is making with its Dynamo inference platform with disaggregated serving.
Dave Salvator, Nvidia’s director for Accelerated Computing Products, said the company’s research has shown disaggregated serving – that is, running different parts of an inference workload on different GPUs – can actually deliver “some pretty significant performance speeds ups.”
For instance, using disaggregated serving with OpenAI’s new GPT-OSS model delivers nearly a 4x increase in tokens per second per user. And using DeepSeek’s R1 model, disaggregated serving delivers a 2.5x increase in throughput in the form of tokens per second per GPU.
“This is another area of ongoing research and ongoing development that we’re working on,” he concluded. “But what we’re seeing here overall is the promise of disaggregated serving is very encouraging in terms of what it can give us, how it gets us more performance out of the same infrastructure.”
AI in telecom isn’t tomorrow’s problem — it’s today’s imperative. If you’re waiting for it to catch up, you’re already behind. Join Fierce Network, Supermicro and Nvidia for an exclusive Accelerating AI Solutions for Telecom event at the Dallas Cowboys HQ. Click here to check your eligibility and register today.